Inference pricing in 2026 tells two contradictory stories at once: per-token prices in free fall, and total inference spend rising almost everywhere. Both are true, and understanding why is the difference between an AI budget you control and one that controls you.
The Price Floor Keeps Dropping
Anchor prices from published rate cards this year:
DeepSeek V4-Flash: $0.14 per 1M input tokens (cache miss), $0.87 output — with a 1M-token context window.
DeepSeek V4-Pro: $0.435 input / $2.19 output — for a 1.6-trillion-parameter (49B active) model.
xAI Grok 4.3: $1.25 input / $2.50 output per 1M tokens.
Anthropic Claude Sonnet 5: launched at introductory $2 / $10, stepping to $3 / $15 — a reminder that launch pricing and steady-state pricing are different numbers.
OpenAI cites a 97% per-token price decline from GPT-4 to its current flagship.
The floor-setter is architecture: sparse mixture-of-experts models activate a small fraction of parameters per token (49B of 1,600B for V4-Pro; 6.5B of 119B for Mistral Small 4), so serving cost scales with active parameters while capability scales with total. That decoupling is why sub-dollar frontier-adjacent inference exists at all.
So Why Are Bills Going Up?
Because volume and token-intensity are outrunning price declines:
Agents multiply tokens per task. A chat answer is a few thousand tokens; an agentic workflow with tool calls, retries, and reasoning traces is tens to hundreds of thousands. Reasoning modes bill their thinking.
Context windows invite context abuse. Million-token windows make it easy to ship whole repositories where a retrieval step would send 2%.
Usage compounds. OpenAI's APIs process more than 15 billion tokens per minute; enterprise AI portfolios add use cases faster than any per-token price falls.
The correct mental model: price-per-token is deflating at tens of percent per year, tokens-per-task is inflating faster in agentic workloads, and the product of the two is your bill.
The Structural Levers (Beyond the Rate Card)
Caching: cached-input pricing makes long static prefixes nearly free; agent stacks with heavy tool schemas live or die by this.
Batch tiers: ~50% off anything asynchronous.
Dedicated capacity: per-instance-hour pricing (Cohere's Model Vault model, and equivalents elsewhere) converts token risk into capacity planning at high utilization.
Self-hosted open weights: at sustained volume, one or two GPUs now serve a competitive enterprise model — the fixed-cost barrier collapsed.
Compatibility-driven arbitrage: OpenAI-compatible APIs (DeepSeek supports OpenAI and Anthropic shapes; xAI is OpenAI-compatible) mean re-routing traffic to the current price/quality winner is a config change. Run the arbitrage quarterly.
Watch the Deprecation Tax
Cheap models carry a hidden cost: lifecycle churn. DeepSeek retired legacy model IDs with hard cutoffs; xAI rebills deprecated slugs at replacement-model rates; Cohere retired embedding models breaking existing calls. Every forced migration costs engineering time and re-evaluation. Price it in — roughly one migration per vendor per year is the current base rate.
Modeling Your Cost Curve
For planning, model three lines per use case: tokens per completed task (measure it — it drifts), tasks per month, and blended price per token after caching/batching/routing. Teams that track all three renegotiate from data; teams that track only the invoice discover agent projects 5× over budget with no idea which line moved.
Internal Link Suggestions
Open vs. Closed AI Models in 2026 → aifn-open-vs-closed-models-2026
The Agentic AI Stack in 2026 → aifn-agentic-ai-stack-2026
LLM Cost Optimization for the Enterprise → aitj-llm-cost-optimization-enterprise
